Google's Secret Weapon Sneaks Out: Gemini Omni Challenges Seedance 2.0

Now, whenever any company releases a video model, it inevitably gets compared to Seedance 2.0.

Recently, Google—previously quiet—appears to be quietly preparing a major move to reclaim the throne of video generation.

This morning, a user accidentally triggered a brand-new video generation model within Gemini—Omni (Omnipotent).

On Gemini’s chat interface, it now prominently displays: “Create with Gemini Omni. Meet our new video generation model. Remix your videos, edit directly in conversation, try rich templates, and more.”

▲ Screenshot of Gemini with unexpected access

However, the user reported that upon reopening the Gemini app later, the Omni feature had disappeared, and the interface reverted to the old version.

From the leaked early test results, Gemini Omni shows impressive performance in certain scenarios—reminiscent of how Nano Banana once dominated image generation.

In two exposed test cases, just as Nano Banana solved text rendering issues, Gemini Omni appears to have addressed two longstanding pain points in AI video generation: text coherence and physical plausibility of character eating.

▲ Video source: X@chetaslua｜Prompt: A professor writes out a mathematical proof for trigonometric identities on a traditional chalkboard, explaining the step he is currently on in the equation.

This 10-second video challenges complex mathematical board writing. The prompt was just one sentence: "A professor writes out a mathematical proof for trigonometric identities on a traditional chalkboard, explaining the step he is currently on in the equation."

Although minor AI-generated artifacts still exist in the frame, the entire video handles the "text" on the chalkboard flawlessly—a major bug in many previous video models. Additionally, the overall realism is striking.

Users reacted with disbelief: “This is too insane.” AI super creator @Azed_ai commented under the leak video: “If this is real output, text consistency is truly unbelievable.”

▲ Video file link: https://gemini.google.com/share/7d5dc678c80a

Some questioned the video’s origin, suggesting it might not have been generated by Gemini Omni. The leaker promptly shared the official Gemini link.

We tested the same prompt on Seedance 2.0—the realism was also impressive, but the board content seemed only loosely related to trigonometric identities, and some text rendering issues remained.

▲ Generated by Seedance 2.0

The other leaked video pays homage to the iconic “Will Smith eating spaghetti” test case—a benchmark used across nearly all video generation models.

▲ Source: Reddit@Zacatac_391｜Prompt: Can you create a scene with two men at a table seaside at an upscale restaurant on outdoor deck seating. They are at a circular table with a nice white table cloth, and all of the fancy accessories, all the spoons forks and knives, fancy napkins, centerpiece. One man is Distinguished: A mature African-American man in his 50s with a short beard and confident posture, wearing a tailored, sophisticated suit, the other is his friend, both approaching the table to eat a plate of spaghetti. In the beginning the men approach the table, exchange brief niceties, and begin to eat the spaghetti calmly. In between bites sharing conversation.

Testers with early access said they were blocked by Google’s content generation safeguards when trying to input the Will Smith spaghetti prompt, so they described a man resembling Will Smith instead.

The resulting video from Gemini Omni surpasses Google’s own Veo 3.1 in audio and visual realism—much closer to real human footage.

Yet some users noticed inconsistencies: when standing, the plate had no spaghetti; once seated, it appeared; then during eating, the noodles vanished completely without being consumed.

We ran the same prompt through Seedance 2.0—the eating behavior looked far more natural than Gemini Omni’s.

▲ Generated by Seedance 2.0｜Prompt: Can you imagine a scene of two men dining at an upscale seaside restaurant on an outdoor deck? They sit at a round table with a pristine white tablecloth, surrounded by elegant utensils—spoons, forks, knives, fancy napkins, and a centerpiece. One man is distinguished: a mature African-American man in his 50s with a short beard and confident posture, wearing a tailored, sophisticated suit. The other is his friend. Both approach the table to eat a plate of spaghetti. Initially, they walk over, exchange brief pleasantries, and begin eating calmly, occasionally chatting between bites.

On Reddit, some questioned whether the tester had used the latest video model. The tester replied he’d used the best available—Kling 3.0.

Another user responded: “Seedance 2.0 is significantly better than Kling 3.0,” so Gemini Omni’s performance doesn’t surprise someone who’s used Seedance 2.0.

On X, others believe Gemini Omni outperforms Seedance 2.0.

▲ Many X users shared these leaks, claiming the results beat Seedance

Even more surprising: Google isn’t just introducing video generation—it’s also launching video editing capabilities with Gemini Omni.

For example, you can directly replace the spaghetti in the original video with a bowl of Thai-style tom yum soup.

You can even edit a watermarked video originally generated by Sora into a clean version without the watermark.

▲ Video source: X@Waguri_Kaoruko8

From the tests, Gemini Omni successfully removed the continuously moving Sora watermark from the original video while maintaining strong consistency with the original footage.

Other users tested Gemini Omni’s performance in anime-style video generation.

Just as the original Nano Banana focused on image editing—enabling users to create various desktop figurines—many now speculate Gemini Omni will follow a similar path: weaker video generation, but superior editing tools. The true video generation prowess may only emerge in Gemini Omni 3.

In the leaked interface, a usage screenshot appeared. It showed that generating just two videos—mathematical board writing and spaghetti eating—consumed 86% of the user’s daily quota under their AI Pro subscription. The user mentioned he’d only used Gemini Flash that day.

AI video remains a costly business—but if the results are compelling enough, there will be plenty of paying customers.

Over recent weeks, AI-generated videos have gone viral online, featuring longer durations, more realistic scenes, dialogue, acting, and storytelling—far surpassing last year’s breakout AI videos.

Marko Slavnic, Runway’s technical artist, posted a pigeon video made using Runway, which supports the Seedance 2.0 model. The video quickly garnered millions of views.

One viewer noted: “The 47-second clip shows zero AI traces or unnatural movements. Disney animation might need to start worrying.”

▲ Video source: X@Markoslavnic

Another trending topic on Weibo—“AI short films mass-producing ‘Love Death & Robots’”—was shared globally and reached 12 million views.

The sharer, PJ Ace, a creator in generative AI video and advertising, and founder of an AI ad agency, stated:

This is one of the best short films I’ve seen in years.

Soon, we won’t call them “AI movies” anymore—we’ll just call them movies.

▲ Video source: X@PJaccetturo

In the original Bilibili video’s comment section, viewers noted its massive popularity—not only did international figure PJ Ace express interest in offering a job, but also an AI+ creative team from Guangzhou reached out.

Praise poured in: “I’d go to the cinema just for this trailer,” “Top-tier work—color, composition, story—all flawless. Only slightly short in length, but would fit seamlessly into any season of *Love Death & Robots*.” “Years ago, this would have taken a top-tier team months and millions to produce…”

▲ Video source: 【AI Original Short Film 'Zombie Janitor' - Remastered Version, China's 'Love Death & Robots'!】https://www.bilibili.com/video/BV1FFRQB2Eqw/

Just as people used to recommend several films each month, now users are starting to compile lists like “Four Must-See AI Shorts This Month.”

Beyond the content criticized as AI slop, the rising quality of AI-generated videos makes it an inevitable trend for them to become mainstream media.

Google's recent unveiling of Gemini Omni may also be aimed at building momentum for next week's Google I/O event. If Gemini Omni is made available at this year's conference, a new chapter in AI video generation will undoubtedly begin.